review-pr-comments

name	review-pr-comments
description	Make an existing PR merge-ready through the complete feedback loop: inventory every bot and human review surface, apply valid same-PR fixes, reply, sync the live target branch, resolve conflicts, push, and recheck until the latest head is stably clean. Use for a full review sweep or merge-ready result, not one selected comment. For the whole commit-to-PR lifecycle use ship-it.
user-invocable	true
argument-hint	[PR number, URL, or 'all' for all open PRs]

Review PR Comments

Dedicated entrypoint for the PR comment loop. The shared policy — mode-split safety, categories, classification context, bot detection, reply format, the inventory + recheck loop, and the clean-stop definition — lives in references/pr-comment-loop.md; API endpoints live in references/github-comment-fetching.md. Load both before acting and follow them; this skill adds only the entrypoint contract, preflight, and flow — it does not restate the policy.

When to run (proactive — don't wait to be asked)

You OWN the review threads on every PR YOU open. Run this loop automatically, scoped to PRs you created in this work (don't retroactively sweep others'), at two moments:

1. ~3–5 min after gh pr create — CodeRabbit, the Codex/ChatGPT connector, and Socket post a few minutes after open, not at open.
2. After every push to that PR branch — bots re-review new commits progressively.

Reporting gate: never report a PR "green / ready / merge-ready / done" until the loop reference's quiet-window stop condition holds. CI-green is not review-clean. The usual misfires: declaring ready at open-time before bots comment, and counting only non-outdated threads while unresolved outdated ones remain.

Inputs

$ARGUMENTS is a PR number, URL, or all. With no argument, detect from the branch: gh pr view --json number -q .number. all is expanded mode — list candidates and confirm before processing (see Mode-Split Safety in the reference).

Preflight

Fail early with a clear blocker, not after edits:

1. gh auth status and gh repo view — CLI authed, repo resolves.
2. git status, branch, remote, push permission. Refuse to work on main unless the user asked.
3. Resolve the target PR, or confirm expanded mode for all.
4. Read repo instruction files (AGENTS.md, CLAUDE.md, or local equivalent).
5. Identify the quality gate (instructions → lockfiles → package scripts → ask).
6. Fetch the PR diff file list before classifying.

Flow

1. Discover every review surface for the PR (commands in github-comment-fetching.md): review comments, review bodies, issue comments, unresolved threads (including isOutdated=true), latest-head checks, and mergeability. Build the inventory per the reference's Inventory Rules.
2. Classify and reply per the reference (it owns categories, required context, bot detection, reply format, thread resolution, and review-body minimization). For each Fix: read the file + surrounding code, apply the smallest change, batch related fixes into conventional commits, run the quality gate — then reply per the reference's policy.
3. Sync and push with the PR's live target branch per the reference's Target-Branch Sync policy. Resolve conflicts, rerun the quality gate, and push.
4. Recheck per the reference's Recheck Loop until its clean-stop condition holds — reviewer surfaces before CI, reset on any new finding.

Summary Report

End with:

PR #123 - Review Complete
Fixed: <count> (each replied with commit SHA on the original thread)
Dismissed: <count> (each replied as false positive on the original thread)
Out-of-scope: <count> (each replied with deferral reason on the original thread)
Needs attention: <thread URLs awaiting user input>
Unreplied: <should be 0 — if not, list thread URLs and reason>
CI:
Merge-ready:
Commits:
Remaining:

Unreplied must be 0 for a clean run (it counts only Fix / False Positive / Out of Scope). Every Needs Discussion thread appears under Needs attention with its URL.